Excel Python

Engineer Excel files from an explicit workbook contract. Preserve package semantics, distinguish stored artifacts from Excel-runtime behavior, mutate the smallest supported region, and prove the result by independent reopening and a semantic comparison.

Boundary and ownership

Required semantics	Owner	Rule
Workbook/package structure, formulas, Tables, names, styles, charts, preservation	excel-python	Inspect OOXML and use openpyxl or XlsxWriter deliberately.
Tabular extraction only	Polars read_excel with Calamine when installed	Use when workbook layout is merely an input boundary.
Joins, aggregation, dtypes, nulls, lazy plans, analytical computation	polars-python	Hand off dataframe decisions; return only for workbook delivery.
Recalculation, refresh, native PDF/rendering, authorized VBA, unsupported Excel objects	installed-Excel runtime workflow	Stop and escalate; this skill never simulates Excel.

Inspect installed APIs before relying on drift-prone behavior. Run python scripts/inspect_excel_env.py from the installed skill and read library selection.

Ordered workflow

1. Recover the workbook contract from the request, workbook, callers, and downstream consumers. State sheet roles, Tables and names, row grain, keys, units, input/output regions, formula-owned columns, editable areas, presentation areas, compatibility, preservation, calculation, and delivery requirements. Prefer cell < range < defined name < Excel Table < workbook contract; use bare addresses only when no stable interface exists. Read workbook contracts.
2. Preserve the source. Resolve extension, output destination, trust boundary, workbook date system, macros, links, and features that must survive.
3. Before editing a valuable workbook, run python scripts/inspect_workbook.py INPUT.xlsx. Treat its JSON as preflight evidence, not proof that openpyxl can preserve every reported part. Unknown or unsupported OOXML plus a preservation requirement blocks a blind round-trip.
4. Choose the operation path: Polars/Calamine for table extraction, openpyxl for supported existing-workbook inspection or minimal mutation, and XlsxWriter or Polars write_excel for new workbooks. Read ingestion, existing workbooks, or generation.
5. Apply explicit type and precision, formula, and untrusted text policies before writing.
6. Write to a temporary sibling, close successfully, then reopen independently. For edits, run python scripts/diff_workbooks.py BEFORE AFTER; classify each difference as required, acceptable serialization variation, or unexpected.
7. Run contract assertions and relevant project tests. For stakeholder output, render or inspect representative sheets because valid OOXML can be unusable. Read semantic validation.

Workbook contract invariants

• A cell format changes display, not stored arithmetic. Identifiers that can exceed 15 significant digits are text. Exact rounding is a value operation, not a number format. Define blank, empty string, NaN, infinity, boolean, date, datetime, timezone, currency, percentage, and basis-point policy.
• Preserve the 1900/1904 date system. Do not transfer raw serials across date systems without conversion and verification.
• Formula text, cached value, and Excel-engine recalculation are three evidence states: FORMULA_TEXT_CHECKED, CACHE_INSPECTED, and RECALCULATED. Claim only states actually established.
• XlsxWriter stores formulas in US-English syntax. It does not calculate them; absent an explicit result it stores zero and requests recalculation on open.
• Externally sourced text is text. Do not allow leading = or URL-like data to become formulas or hyperlinks implicitly; use write_string() or disable strings_to_formulas and strings_to_urls for text-first exports.
• For .xlsm, use keep_vba=True, preserve the macro-enabled extension, never execute VBA, and compare the VBA project hash before and after.
• Never execute macros, refresh connections, claim remote link freshness, or claim Excel recalculation from an openpyxl/XlsxWriter save.

Preservation and scale gates

Inspect ZIP parts and relationships for VBA, drawings/shapes, charts, chartsheets, pivots and caches, slicers, external links, connections/Power Query artifacts, custom XML, embeddings, ActiveX/forms, comments/threaded comments, media, custom properties, and unknown relationships. If a required feature is unsupported or preservation is unknown, do not save with openpyxl; retain the original and require an authorized Excel-native path. Read OOXML preservation.

Fail before exceeding 1,048,576 rows or 16,384 columns per worksheet. Prefer a summarized Excel view with Parquet/DuckDB/Polars for oversized analytical data. XlsxWriter constant_memory requires sequential rows and disables Tables and merged ranges; choose features before memory mode. openpyxl read-only/write-only modes also omit normal APIs. Read performance and limits.

Generated workbook quality

Keep machine-readable raw data separate from presentation sheets. Use named Tables, visible units, centralized number formats, filters, useful freeze panes, clear input/formula/output styling, stable chart sources, labeled axes, sensible widths, and no merged cells inside data regions. Use hidden sheets only when their role is documented in the contract. Add an optional _meta or About section when provenance matters: generator/schema version, generation and as-of timestamps, source identifiers or hashes, parameters, and code revision. Read presentation.

Completion

Do not declare completion until the source remains intact; preflight risks are resolved; the output reopens; sheet/name/Table/formula/type/date/visibility and selected style/value contracts pass; required VBA and high-risk package parts are unchanged; the semantic diff contains no unexplained differences; visual review ran when presentation is contractual; and recalculation/runtime evidence is reported separately. Use evaluated recipes only for their stated conditions.

References

• Library selection and adjacent-skill routing
• Workbook contracts
• Tabular ingestion
• Types and precision
• Formula semantics
• Inspection
• OOXML preservation
• Existing-workbook mutation
• Generation
• Presentation and provenance
• Performance and Excel limits
• Untrusted text and security
• Semantic validation and testing
• Evaluated core recipes